Class imbalance learning for network fault prediction
The quality of internet connection is important in our daily life as people are becoming more dependent on it. To reduce internet connection problems, internet service providers (ISP) are proactively predicting and identifying network faults before it occurs. One of the network fault prediction meth...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| التنسيق: | أطروحة |
| منشور في: |
2020
|
| الموضوعات: | |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | The quality of internet connection is important in our daily life as people are becoming more dependent on it. To reduce internet connection problems, internet service providers (ISP) are proactively predicting and identifying network faults before it occurs. One of the network fault prediction methods is by classifying incoming network data from Remote Authentication Dial-In User Service (RADIUS) server. Although RADIUS data provide useful accounting information such as total download, cause of termination, etc., it does not provide any indication whether the network is faulty or not. A customer trouble ticket (CTT) is often raised by network user when he/she faces an internet connection problem. In this thesis, a method is proposed to annotate network data by correlating RADIUS data and CTT using temporal interval relations. Once RADIUS data are labelled, a network fault prediction model can be trained. However, performance of the prediction model is greatly affected by class imbalance problem, where the number of faulty cases is relatively rare. One of the challenges in addressing class imbalance problems for network fault prediction is the overlapping between faulty and non-faulty classes, which mainly due to unreported network issues. This thesis proposes a method based on edited nearest neighbours (ENN), Tomek-link and SMOTE to address the class overlapping and class imbalance problems. The findings showed that RADIUS data annotated with our method achieved a higher recall and precision rate. The F1-score and AUC results showed that our proposed method has greatly improved class imbalance learning for network fault prediction. |
|---|